DeepFeature

DeepFeature applies convolutional neural networks to transformed high-dimensional omics data to perform feature selection and predictive modeling for identification of biologically relevant genes and pathways.

Key Features:

• Omics-to-image transformation: Converts high-dimensional omics samples into an image-like format to enable CNN analysis of non-image data.
• CNN-based feature selection and prediction: Uses convolutional neural networks to select discriminative features and generate predictive models from transformed data.
• Class activation maps: Employs class activation maps to localize and prioritize genes or elements that contribute to model predictions.
• Snowfall compression algorithm: Applies the Snowfall compression algorithm to optimize the pixel framework and accommodate more elements in the transformed images.
• Region accumulation and element decoder: Uses region accumulation and an element decoder to extract meaningful biological features from class activation maps.

Scientific Applications:

• Cancer type prediction: Predicts cancer types from high-dimensional omics profiles using CNN-derived features.
• Pathway and process discovery: Identifies significant biological pathways and processes by mapping selected features to genes and pathways.
• Mechanistic insight in complex datasets: Facilitates discovery of candidate causal mechanisms within complex biomedical and omics datasets.

Methodology:

Transforms omics data into image-like format, processes transformed data with CNNs using class activation maps, applies the Snowfall compression algorithm to optimize the pixel framework, and uses region accumulation and an element decoder to extract important genes and pathways.